Data-Efficient Learning-Based Iterative Optimization Method With Time-Varying Prediction Horizon for Multiagent Collaboration

Abstract

Learning-based strategy can be well integrated with model-based optimal control to facilitate cooperative multiagent control through the Internet of Things (IoT). In this work, we propose a data-efficient learning-based iterative optimization method with time-varying prediction horizon (TV-LIO) for multiagent collaboration. Our method builds a multiagent optimization problem by introducing a time-domain guided terminal set and an approximated general cost. We collect the historical agent states at previous iterations as a dataset to reconstruct the general cost and the terminal set iteratively, forming closed-loop data-efficient learning. We consider the influence of the predictive time domain on the optimality and feasibility of the optimization problem and design a time-domain recursive updating mechanism to determine the optimal predictive horizon for each agent at the epoch. The continuous feasibility, stability, and recursive convergence of the proposed method are analyzed theoretically. Unlike the traditional optimization approaches that rely on a preplaned reference path, the proposed method integrates the trajectory planning and tracking control for multiple agents. After several iterations, the general cost of the optimization problem monotonically decreases and the optimal states are finally obtained. The proposed approach is validated and the results demonstrate that our approach can obtain the optimal-cost strategy and trajectories with optimizing time domains for the multiagent system.